Seismic Signal Classification Based On EMD-VMD-LSTM Integrated Model And Entropy Component

Earthquakes may incur heavy damage.Researchers in the field of seismology generally comprehend the course of occurrence and transmission of earthquakes by monitoring seismic signals.However,the origination of seismic signals is rather complex,which may not only be originated from natural seismic events,but also being originated from non-natural seismic events such as artificial blasting,chemical explosion,underground nuclear explosion and so on.Due to the limitations of the conditions,the non-natural seismic events involved in the relevant researches in this thesis only include the artificial blasting events,and the other main reason for considering the artificial blasting events is that such engineering activities such as artificial blasting are also the more frequent being observed in the seismic observatory network.In order to solve the discrimination problem of seismic signals originating from natural earthquakes and man-made explosion,this thesis proposes a model method which combining empirical modal decomposition(EMD),variable modal decomposition(VMD)and long-short-term memory network(LSTM);makes an attempt to the combination of time-frequency graph and deep learning methods,and furthermore proposes the entropy component features for seismic signals classification.The main research contents include:(1)A model method based on EMD-VMD-LSTM is proposed.The two modern signal processing methods-empirical modal decomposition(EMD)and variable modal decomposition(VMD)are combining to be applied to seismic signals processing,and also making use of long-short-term memory network(LSTM)'s better ability to process time series,this thesis proposes the EMD-VMD-LSTM model,which completing the classification and identification of seismic signals.(2)The entropy component is used as a feature to classify seismic signal.Seismic waveforms are decomposed by wavelet packets.A data matrix of wavelet packet coefficients can be obtained from each seismic waveform.The singular value decomposition of the rectangle is carried out to the data matrix.The larger singular values in the front are the main singular values and are kept for calculating entropy components by Shannon entropy method.(3)Making attempt to conduct seismic signals classification experiments by combining the time-frequency graph with deep learning method.Firstly,by using S-transform,wavelet transform,short-term Fourier transform,conducting various the time-frequency transforms to seismic signal,the time-frequency transform images of the seismic signal are acquired.And then by utilizing VGG16,Resnet50 deep learning method,some classifiable features can be extracted from these time-frequency images.(4)Established a software system for seismic signal processing.The software provides good interaction and better management of seismic signal and events.A wide variety of feature extraction algorithms and classification recognition algorithms are integrated into this software system,such as the capability to compute signal energy values,to show waveform in real time.This accomplishment has consumed considerable time in the endeavor of this study,but it provides the necessary basis for the successful carryout of the study.This thesis proposes the EMD-VMD-LSTM classification recognition model which is an integration model by combining the classical methods.This method not only maintains the overall observation of the seismic signal,but also further improves the classification recognition rate.By combining the classic time frequency analysis method and deep learning method can eliminate the step of manually extracting features,and convert the signal into an image,which can more intuitively analyze the signal.The combination of time-frequency graph and deep learning has achieved the correct recognition rate of 87.82% with a single waveform signal as the classification recognition unit,which shows that the deep learning method implying potential wide application prospect in seismic signal processing.Inspired by the classical entropy characteristic method,the entropy component value is used as a characteristic method,the recognition unit of the event is used,the correct recognition rate of earthquake event is 96.67%,and the correct recognition rate of blasting event is 94.29%,which shows that the entropy component value is a highly reliable seismic signal classification recognition feature.